MACHINE LEARNING-BASED FOURIER AMPLITUDE SPECTRA MODELING IN TÜRKİYE
Abd Al-Rahman Al-Asadi

(Thesis Supervisor: Prof. Dr. Gülüm Tanırcan)

ABSTRACT

This study develops a Türkiye-specific machine learning framework for predicting the Fourier Amplitude Spectrum (FAS) of strong ground motion. The proposed ANN model predicts log10 effective amplitude spectrum (EAS) at 19 discrete frequencies using a Türkiye-specific earthquake catalog comprising 752 shallow crustal events from the active tectonic setting of the Anatolian region.

After evaluating the effect of dataset composition, the final model is trained using the moderate-to-large event subset, consisting of 125 events with Mw ≥ 5, to improve the model’s relevance to engineering-significant ground motions. The input features are Mw, Rjb, Vs30, and style of faulting. To prevent event-level data leakage, the dataset is divided into 70% training, 15% validation, and 15% testing using an event-isolated controlled split. The final framework incorporates normalization, a Wide vector-output ANN architecture, and a mixed-effects-inspired loss function.

Model performance is evaluated using MSE, MAE, RMSE, R2, and R, in addition to residual analysis and comparison with two conventional parametric models. The results show that the ANN captures frequency-dependent behavior, performs strongest in the low-to-mid frequency range, remains competitive with the regional conventional model within the common frequency range, and outperforms the external benchmark on the Türkiye-specific testing dataset. Residual and scenario-based analyses indicate that near-source high-frequency distance scaling remains the main limitation of the developed model. Overall, the study shows that an event-aware ANN can provide a promising complementary framework for regional FAS/EAS ground-motion modeling in Türkiye.